Time-varying causal interactions in the functional brain networks: Modeling and Validation

1750 

Abstract Submission 

Nan Xu¹, Xiaodi Zhang¹, Wen-Ju Pan², Jeremy Smith², Jason Allen³, Vince Calhoun⁴, Shella Keilholz¹

¹Georgia Institute of Technology, Atlanta, GA, ²Emory University, Atlanta, GA, ³Indiana University School of Medicine, Indianapolis, IN, ⁴GSU/GATech/Emory, Decatur, GA

Nan Xu  
Georgia Institute of Technology
Atlanta, GA

Xiaodi Zhang  
Georgia Institute of Technology
Atlanta, GA

Wen-Ju Pan  
Emory University
Atlanta, GA

Jeremy Smith  
Emory University
Atlanta, GA

Jason Allen  
Indiana University School of Medicine
Indianapolis, IN

Vince Calhoun  
GSU/GATech/Emory
Decatur, GA

Shella Keilholz  
Georgia Institute of Technology
Atlanta, GA

The fMRI BOLD brain dynamic processes have served as a sensitive indicator for various brain disorders [1]. Most work in characterizing brain dynamics evaluates temporal variability among functional brain networks e.g., [2], [3]. There has been less focus on dynamic analysis to capture the brain's causal interactions [4], [5], which may play a key role in understanding clinical outcomes in conditions like post-concussive visual motion sensitivity (PCVMS). To overcome this, we propose a new computational method to reveal time-varying causal dynamics in brain networks. The model's reliability was validated through concurrent LFP-BOLD measurements in rodents, assessing if causal interactions stem from deeper neuronal layers. Further validation was achieved by observing brain state shifts in the vestibular networks of PCVMS patients.

The model integrates the sliding-window technique with a validated static causal model prediction correlation [6], [7] to analyze time-varying causal interactions in brain networks. It operates in two main steps within each window [t, t+L] (Fig 1): (1) using a linear time-invariant model to predict output y from input x and determine the duration of information transfer D_yx[t], and (2) estimating directed functional connectivity strength (ρ_yx[t]=corr(ý[t], y[t]), where ý[t] is the prediction of y[t] from x[t] in step (1)). This method called SWpC dynamically predicts connectivity strength and duration using inputs from current and previous windows, distinguishing between forward and backward connections (ρ_yx[t]≠ρ_xy[t] and D_yx[t] ≠D_xy[t] for any t).

SWpC was first validated with 22 concurrent LFP-BOLD scans (500s/scan) from two regions of the rat somatosensory cortex (S1L and S1R) under 1-2% dexmedetomidine [8], assessing the reliability of signal strength and duration. Band-limited power (BLP) signals across six frequency bands (δ, θ, α, low β, high β, and γ) was calculated from downsampled LFP using a 2TR-sliding window [9], matching BOLD's sampling rate (TR=0.5s). SWpC was then applied to both BOLD and BLP signals of S1L and S1R, contrasting with sliding window correlation (SWC) results. Following [10], a 50s sliding window was used in SWpC and SWC analysis. Second, SWpC assessed alterations in brain states, both in strength and duration, using fMRI recordings (420s/scan, TR=0.7s) of the vestibular networks [11] in patients with varying severity of PCVMS (37 healthy controls, 25 subacute, 15 chronic). Following previous human study [2], a 44s sliding window was implemented and brain states identified by strength and duration were obtained by k-mean clustering.

First, strength and duration of SWpC were reliably predicted in both BLP and BOLD signals from S1L and S1R, exhibiting symmetries between the two regions as their directional asymmetry is consistently < scan variability (Fig 2A). SWpC strength reflects LFP-BOLD correlations akin to those uncovered by SWC [9], showing high BLP-BOLD correlates within the θ, low β, and upper bands, while SWpC duration lacks significant LFP-BOLD correlation (Fig 2BC). Second, four SWpC brain states in strength and duration were identified in PCVMS patients. Severity influenced time spent in strength-based states, with subacute similar to controls and chronic showing more deviation, consistent with known patterns of the disorder. However, less variation was noted in duration-based states (Fig 2D).

A computational framework was developed to reliably estimate the dynamic strength and duration of information transfer in neural and BOLD signals. BOLD strength estimates reflect causal dynamics in LFP at θ, low β and upper bands , but BOLD estimates in temporal influence do not tie to neural activity. This aligns with prior findings of LFP-BOLD correlates using SWC [9] and expands such correlates to include causal factors. Furthermore, SWpC estimations offer fresh perspectives on the dynamic nature of neurological disorders like PCVMS.

Connectivity (eg. functional, effective, structural) ²

fMRI Connectivity and Network Modeling ¹

Methods Development

Multivariate Approaches

ADULTS

ANIMAL STUDIES

Computational Neuroscience

Data analysis

DISORDERS

FUNCTIONAL MRI

Informatics

Modeling

Multivariate

Statistical Methods